speech_siamese.py

@@ -4,9 +4,8 @@ import numpy as np
 from speech_data import speech_model_data
 from keras.models import Model,load_model
 from keras.layers import Input, Dense, Dropout, LSTM, Lambda, Concatenate
-from keras.losses import categorical_crossentropy
-# from keras.losses import binary_crossentropy
-from keras.utils import to_categorical
+# from keras.losses import categorical_crossentropy
+from keras.losses import binary_crossentropy
 # from keras.utils.np_utils import to_categorical
 from keras.optimizers import RMSprop
 from keras.callbacks import TensorBoard, ModelCheckpoint
@@ -31,14 +30,15 @@ def contrastive_loss(y_true, y_pred):
     return K.mean(y_true * K.square(y_pred) +
                   (1 - y_true) * K.square(K.maximum(1 - y_pred, 0)))
 
+
 def create_base_rnn_network(input_dim):
     '''Base network to be shared (eq. to feature extraction).
     '''
     inp = Input(shape=input_dim)
-    ls1 = LSTM(256, return_sequences=True)(inp)
-    ls2 = LSTM(128, return_sequences=True)(ls1)
+    ls1 = LSTM(1024, return_sequences=True)(inp)
+    ls2 = LSTM(512, return_sequences=True)(ls1)
     # ls3 = LSTM(32, return_sequences=True)(ls2)
-    ls4 = LSTM(64)(ls2)
+    ls4 = LSTM(32)(ls2)
     return Model(inp, ls4)
 
 
@@ -52,15 +52,15 @@ def compute_accuracy(y_true, y_pred):
 def accuracy(y_true, y_pred):
     '''Compute classification accuracy with a fixed threshold on distances.
     '''
-    return K.mean(K.equal(y_true, K.cast(y_pred > 0.5, y_true.dtype)))
+    return K.mean(K.equal(y_true, K.cast(y_pred < 0.5, y_true.dtype)))
 
 def dense_classifier(processed):
     conc_proc = Concatenate()(processed)
-    d1 = Dense(16, activation='relu')(conc_proc)
-    # dr1 = Dropout(0.1)(d1)
-    d2 = Dense(8, activation='relu')(d1)
+    d1 = Dense(8, activation='relu')(conc_proc)
+    dr1 = Dropout(0.1)(d1)
+    # d2 = Dense(8, activation='relu')(dr1)
     # dr2 = Dropout(0.1)(d2)
-    return Dense(2, activation='softmax')(d2)
+    return Dense(1, activation='sigmoid')(dr1)
 
 def siamese_model(input_dim):
     # input_dim = (15, 1654)
@@ -80,9 +80,7 @@ def siamese_model(input_dim):
 
 def train_siamese():
     # the data, shuffled and split between train and test sets
-    tr_pairs, te_pairs, tr_y_e, te_y_e = speech_model_data()
-    tr_y = to_categorical(tr_y_e, num_classes=2)
-    te_y = to_categorical(te_y_e, num_classes=2)
+    tr_pairs, te_pairs, tr_y, te_y = speech_model_data()
     input_dim = (tr_pairs.shape[2], tr_pairs.shape[3])
 
     model = siamese_model(input_dim)
@@ -98,7 +96,7 @@ def train_siamese():
         embeddings_layer_names=None,
         embeddings_metadata=None)
     cp_file_fmt = './models/siamese_speech_model-{epoch:02d}-epoch-{val_loss:0.2f}\
--acc.h5'
+    -acc.h5'
 
     cp_cb = ModelCheckpoint(
         cp_file_fmt,
@@ -110,7 +108,7 @@ def train_siamese():
         period=1)
     # train
     rms = RMSprop(lr=0.001)
-    model.compile(loss=categorical_crossentropy, optimizer=rms, metrics=[accuracy])
+    model.compile(loss=binary_crossentropy, optimizer=rms, metrics=[accuracy])
     model.fit(
         [tr_pairs[:, 0], tr_pairs[:, 1]],
         tr_y,

test_siamese.py

@@ -5,7 +5,6 @@ import speech_data
 reload(speech_data)
 from speech_data import create_test_pair,get_word_pairs_data
 import numpy as np
-from keras.utils import to_categorical
 
 model = siamese_model((15, 1654))
 model.load_weights('./models/siamese_speech_model-final.h5')
@@ -16,9 +15,8 @@ def predict_recording_with(m,sample_size=15):
     inp = create_test_pair(spec1,spec2,sample_size)
     return m.predict([inp[:, 0], inp[:, 1]])
 
-while(True):
-    print(predict_recording_with(model))
+predict_recording_with(model)
 
-# sunflower_data,sunflower_result = get_word_pairs_data('sweater',15)
-# print(np.argmax(model.predict([sunflower_data[:, 0], sunflower_data[:, 1]]),axis=1))
-# print(sunflower_result)
+sunflower_data,sunflower_result = get_word_pairs_data('sunflowers',15)
+sunflower_result
+model.predict([sunflower_data[:, 0], sunflower_data[:, 1]]) < 0.5

tts_samplegen.py

@@ -11,7 +11,7 @@ import progressbar
 
 from generate_similar import similar_phoneme,similar_word
 
-OUTPUT_NAME = 'rand_edu'
+OUTPUT_NAME = 'story_words'
 dest_dir = os.path.abspath('.') + '/outputs/' + OUTPUT_NAME + '/'
 dest_file = './outputs/' + OUTPUT_NAME + '.csv'
 
@@ -95,7 +95,8 @@ class SynthVariant(object):
         self.name = voice
         self.lang = lang
         self.phoneme_capable = self.is_phoneme_capable()
-
+        if self.phoneme_capable:
+            create_dir(dest_dir + self.name + '/' + str(self.rate))
 
     def __repr__(self):
         return 'Synthesizer[{} - {}]'.format(self.name, self.rate)
@@ -127,10 +128,6 @@ class SynthVariant(object):
         cli_gen_audio(phon_cmd, self.rate, self.name, d_path)
         return SynthFile(word, phoneme, r_path, self.name, self.lang, self.rate, variant)
 
-    def create_synth_dirs(self):
-        if self.phoneme_capable:
-            create_dir(dest_dir + self.name + '/' + str(self.rate))
-
     @staticmethod
     def voices_for_lang(lang):
         voices_installed = NSSpeechSynthesizer.availableVoices()
@@ -171,10 +168,9 @@ def synth_generator():
 
     def synth_for_words(words, writer):
         prog_title = "Synthesizing {} words : ".format(len(words))
+        (update, prog) = prog_bar(prog_title)
         for s in voice_synths:
-            s.create_synth_dirs()
             for v in ['low', 'medium', 'high']:
-                (update, prog) = prog_bar(prog_title)
                 for w in prog(words):
                     update('"{}" with {} variant ({})'.format(w, s, v))
                     synthed = s.generate_audio(w, v)
@@ -213,24 +209,13 @@ def synth_logger(fname, csv=False):
     else:
         return json_writer, close_file
 
-def generate_audio_for_text_list(text_list):
-    (writer, closer) = synth_logger(dest_file, csv=True)
-    synth_for_words = synth_generator()
-    try:
-        synth_for_words(text_list, writer)
-    except:
-        import traceback
-        import sys
-        traceback.print_exc(file=sys.stdout)
-        pass
-    closer()
 
 def generate_audio_for_stories():
-    # story_file = './inputs/all_stories_hs.json'
-    story_file = './inputs/all_stories.json'
+    story_file = './inputs/all_stories_hs.json'
+    # story_file = './inputs/all_stories.json'
     stories_data = json.load(open(story_file))
-    # word_list = [t[0] for i in stories_data.values() for t in i]
-    word_list = [i for g in stories_data.values() for i in g]
+    word_list = [t[0] for i in stories_data.values() for t in i]
+    # word_list = [i for g in stories_data.values() for i in g]
     (writer, closer) = synth_logger(dest_file, csv=True)
     synth_for_words = synth_generator()
     try:
@@ -243,5 +228,4 @@ def generate_audio_for_stories():
     closer()
 
 if __name__ == '__main__':
-    generate_audio_for_text_list(['random','education'])
-    # generate_audio_for_stories()
+    generate_audio_for_stories()